Glass is a highly recyclable material that provides substantial environmental benefits, including savings in raw materials and energy as well as a reduction in CO2 emissions. To ensure the production of high-quality secondary raw materials, container glass from municipal waste separate collection must be accurately separated by color in recycling plants, where only minimal color mixing is tolerated. Color sorting is therefore a key step in glass recycling, as it directly affects both the quality and the market value of recycled cullet. Given the increasingly stringent color quality requirements for recycled glass and the high fraction of cullet used in container glass, advanced technological solutions are needed to improve sorting accuracy. In this study, a visible–near-infrared (VIS-NIR: 400–1000 nm) hyperspectral imaging (HSI) approach integrated with machine learning (ML) is proposed for the automated classification of post-consumer glass fragments from bottles and jars into five color categories: brown, dark green, light green, half-white and white. A hierarchical Partial Least Squares-Discriminant Analysis (PLS-DA) model combined with an object-based analysis strategy was developed to optimize color recognition. The proposed system achieved sensitivity and specificity values between 0.910 and 1.000, demonstrating excellent robustness and predictive capability. Validation on independent datasets confirmed the model’s reliability, with all color glass fragments correctly classified at the object level. The results highlight the potential of HSI-ML systems to enhance color sorting accuracy and process efficiency in recycling plants, contributing to improved material recovery and the advancement of sustainable, circular glass production.
Bonifazi et al. (Sun,) studied this question.